Optimizing a Fine-Tuning Data Pipeline

A development team is fine-tuning a language model to act as a programming assistant. Their initial training data consists of thousands of simple instruction-code pairs, such as a prompt asking for a function to add two numbers and the corresponding correct code. After training, they observe that the model performs well on simple, one-step tasks but consistently fails to generate correct code for complex problems that require breaking the problem down into multiple logical steps. Which of the following advanced data construction strategies is most likely to address this specific performance issue?

Targeted SFT Data Curation for Stylistic Control